A label classification method based on big data analysis

Abstract

This invention discloses a label classification method based on big data analysis, comprising: S1, collecting and cleaning multi-source heterogeneous raw big data to construct an initial sample pool; S2, mapping a high-dimensional feature space and mining implicit associations to construct a heterogeneous data association map; S3, inputting a dual-stream coupled network to extract macroscopic and microscopic features, and outputting feature vectors after fusion; S4, introducing a data gravity field dynamic clustering algorithm to generate label cluster centers and initial results; S5, using a challenge validation set for calibration and calculating confidence weights to correct the classification probability distribution; S6, combining improved ADWIN and DDM algorithms to monitor distribution drift and initiate dynamic evolution to update the map and cluster centers; S7, matching the optimal label path to generate the final classification result. This invention achieves deep fusion representation of multi-source heterogeneous data, adapts to changes in data distribution in real time through a dynamic evolution mechanism, and improves the accuracy and real-time response capability of label classification.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and big data analysis technology, and in particular to a tag classification method based on big data analysis. Background Technology

[0002] Multimodal data labeling and classification methods based on deep learning and graph computing have been widely applied in recent years in fields such as smart city management, business intelligence analysis, and industrial internet monitoring due to their ability to automatically process and understand massive and complex data. They have become an important technical means to realize the value and structure of data resources. However, in practical applications, facing the continuously growing and diverse data streams, labeling scenarios face many challenges, such as dynamic changes in data distribution, difficulties in integrating multi-source information, and lagging updates to classification standards. The actual deployment effectiveness of existing technologies is still constrained by many factors.

[0003] Most current label classification methods rely on feature extraction from a single type of data, making it difficult to fully utilize the complementary advantages of different dimensions of information such as text descriptions, images, and videos. This results in a lack of comprehensive feature modeling of data content. Some systems only use static threshold partitioning or traditional clustering algorithms to process data, ignoring the feature differences between long-term historical data and instantaneous real-time data, thus limiting the adaptability of classification models to dynamic environments. At the same time, the classification logic process lacks an effective multi-source information interaction mechanism and robust clustering evaluation strategy, making it difficult to accurately capture implicit relationships between data. This leads to biases in the labels generated in complex scenarios, affecting the accuracy and usability of the label results.

[0004] Furthermore, existing label classification technologies often lack dynamic adjustment and evolution mechanisms when faced with changes in data distribution over time. They fail to monitor the distribution characteristics of data in real time and correct the classification model in a timely manner, resulting in the model being unable to maintain stable performance when facing changes in the data environment. This makes it difficult to adapt to the continuous evolution of data scenarios and seriously affects the practical value and long-term stability of the model in real and complex data environments.

[0005] Therefore, how to provide a label classification method based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a label classification method based on big data analysis. This invention fully integrates key steps such as heterogeneous graph construction, dual-stream coupled feature extraction networks, gravitational field dynamic clustering, confidence calibration, and the introduction of an improved ADWIN algorithm with a DDM algorithm early warning hierarchy. It constructs an intelligent label recognition and evolution process with unified mapping of multi-source heterogeneous data, bidirectional interaction of long-term and short-term features, dynamic adaptation of cluster centers, and hierarchical response to concept drift. This achieves accurate label matching and real-time updates in complex data environments. This invention possesses advantages such as multi-dimensional feature extraction fusion, strong anti-perturbation capability, timely evolution response, and strong adaptive weight adjustment. It can significantly improve classification stability, label path accuracy, and long-term model availability during data distribution drift, thereby effectively solving problems such as single feature extraction, lagging model evolution, and fixed classification results in existing methods.

[0007] A tag classification method based on big data analysis according to an embodiment of the present invention includes the following steps: S1. Simultaneously collect raw big data from multiple heterogeneous sources, and perform cleaning, noise reduction, and missing value completion to construct an initial sample pool; S2. Map the initial sample pool to a unified high-dimensional feature space, mine the implicit associations between data nodes, and construct a heterogeneous data association graph; S3. Input the heterogeneous data association map into the dual-stream coupled feature extraction network to extract the macro-level teacher flow with long-term semantic dependence and capture the micro-level student flow with transient mutation features. Through a bidirectional interactive fusion mechanism, the heterogeneous data association map is used to extract and fuse features and output a fused feature vector. S4. Introduce the fused feature vector into the dynamic clustering algorithm based on the data gravitational field. Treat the data samples as spatial particles with mass and density, calculate the gravitational potential energy distribution in the high-dimensional feature space, search for the depth and center of the gravitational potential well, and generate the label clustering center and the initial classification result. S5. Read the initial classification results, initialize a classification decision set and a challenge verification set. In each round of calibration, the challenge verification set generates perturbation data and calculates the stability index until the stability index reaches the preset convergence threshold or the maximum number of rounds is reached. Calculate the confidence weight, perform weighted correction, and output the corrected classification probability distribution and confidence score. S6. Based on the corrected classification probability distribution and confidence score, the improved ADWIN algorithm is used to monitor the distribution characteristics in real time. The early warning level of the DDM algorithm is introduced. When the statistical distance exceeds the warning threshold set by the DDM algorithm, the pre-evolution mechanism is started. When the drift judgment threshold set by the improved ADWIN algorithm is exceeded, the dynamic evolution command is triggered to incrementally update the heterogeneous data association map and label cluster center. S7. Based on the incrementally updated heterogeneous data association map and label cluster centers, match the optimal label path from the preset multi-level label library, generate the final structured label classification result and output it.

[0008] Optionally, S1 specifically includes: S11. Simultaneously start the text data reading process and the image and video acquisition process through the preset synchronous triggering mechanism to collect multi-source heterogeneous raw big data, and respectively obtain the text data stream containing business records and the image and video data containing real-time scenes. S12. Scan each data record in the multi-source heterogeneous raw big data, use regular expression matching to identify non-standard characters and abnormal values, perform deletion operations on the identified non-standard characters and abnormal values, detect whether there are empty values ​​in the key fields of the data record, and when there are empty values, retrieve the statistical feature values ​​in the historical data and use the statistical feature values ​​to fill the empty values. S14. Map the text data and image / video data that have been cleaned, denoised, and had missing value imputation operations to a unified storage structure for aggregation, and construct an initial sample pool.

[0009] Optionally, S2 specifically includes: S21. Extract the pixel matrix of the image and video data in the initial sample pool and expand it into a numerical sequence. Use the bag-of-words model to convert the text data into word frequency vectors. Input the numerical sequence and word frequency vectors into a preset fully connected layer for operation and map them into a vector space with a dimension of 128 to generate node feature vectors. S22. Traverse all node feature vectors, multiply the corresponding dimension values ​​of any two node feature vectors and sum them to obtain the inner product, calculate the product of the moduli of the two node feature vectors, divide the inner product by the product to obtain the cosine value of the included angle, and when the cosine value of the included angle is greater than the preset similarity threshold, establish an undirected edge between the two nodes. S23. Starting from any node feature vector, traverse the network structure to obtain the neighbor node sets of two adjacent nodes and denote them as set A and set B respectively. Calculate the number of elements in the intersection of set A and set B and the number of elements in the union. Divide the number of elements in the intersection by the number of elements in the union to obtain the Jaccard coefficient. When the Jaccard coefficient is greater than a preset coefficient threshold, it is determined that there is a hidden association between the adjacent nodes. S24. Read the current horizontal and vertical coordinate values ​​of every two nodes in the heterogeneous data association map in the two-dimensional coordinate system, calculate the straight-line distance between the two nodes, calculate the repulsion force value by dividing the preset repulsion constant by the square of the straight-line distance, divide by the node mass to obtain the acceleration value, multiply the acceleration value by the preset time step to obtain the velocity value, and superimpose it on the current horizontal and vertical coordinates along the opposite direction of the line connecting the two nodes, and adjust the coordinate positions of the two nodes to make the nodes move away from each other. S25. Read the current horizontal and vertical coordinates of each pair of nodes connected by undirected edges in the heterogeneous data association graph in the two-dimensional coordinate system, calculate the straight-line distance between the two nodes, subtract the preset ideal side length distance to obtain the length difference, multiply the preset elastic coefficient by the length difference to calculate the tension value, divide by the node mass to obtain the acceleration value, multiply the acceleration value by the preset time step to obtain the velocity value, and superimpose it on the current horizontal and vertical coordinates along the direction of the line connecting the two nodes, and adjust the coordinate positions of the two nodes to make the nodes closer to each other. S26. The calculated repulsive force value and tensile force value are vector-superimposed in the horizontal and vertical coordinate directions to obtain the total displacement value. The total displacement value is added to the current horizontal and vertical coordinates of the node to obtain the updated coordinate position. The above calculation and update operation is repeated until the change in the node coordinate position is less than the set error threshold, and a heterogeneous data association map is constructed.

[0010] Optionally, S3 specifically includes: S31. Input the heterogeneous data association map into the dual-stream coupled feature extraction network, read the data records in the initial sample pool, and divide the data records with timestamps earlier than the preset segmentation time into historical big data according to the timestamps of the data records, and divide the data records with timestamps later than the preset segmentation time into real-time data streams. S32. Configure a deep network structure containing three preset convolutional layers and two preset fully connected layers. Input historical big data into the macro teacher flow. In each convolutional layer, use preset convolutional kernels to perform sliding window convolution operations on the feature map of historical big data to extract long-term semantic dependency features in historical big data and generate macro feature maps. S33. Configure a lightweight network structure containing a Transformer encoder, read the image pixel matrix of the real-time data stream, and cut it into image blocks of a preset size in order from left to right and from top to bottom. Straighten the pixel values ​​in each image block into a one-dimensional column vector and arrange them in order to generate a vector sequence. Add the vector sequence to the preset learnable position encoding vector to obtain the position embedding vector. S34. Multiply the position embedding vector by three preset weight matrices to obtain the query vector, key vector, and value vector. Calculate the product of the query vector and the transpose of the key vector and divide by the square root of the vector dimension to obtain the attention score. Perform an exponential operation on the attention score and sum it row-wise to obtain the normalized denominator. Divide the attention score by the normalized denominator to obtain the autocorrelation weight matrix. Perform matrix multiplication on the autocorrelation weight matrix and the value vector to obtain the weighted feature vector. Pass the weighted feature vector through a preset feedforward neural network layer for linear transformation and nonlinear activation to output the micro-feature map. S35. Execute the two-way interactive fusion mechanism, read the average pixel value of each channel in the macro feature map, calculate the product of the average pixel value and the preset guidance coefficient, and perform an addition operation on the calculated product value and the real-time data stream. The macro feature map extracted from historical big data continues to guide the real-time data stream to focus on long-term value features. S36. Read feature points in the micro feature map whose gradient magnitude is greater than the preset mutation threshold, extract the position index and feature value of the feature points in the micro feature map, locate the corresponding position in the macro feature map corresponding to the historical big data according to the position index, add the feature value to the feature value of the corresponding position in the macro feature map, and continuously correct the macro feature map by the feedback of instantaneous abnormal features captured by the real-time data stream. S37. The macroscopic feature map and microscopic feature map after guided correction are concatenated and operated on in the channel dimension. The concatenated feature map is input into the final 1x1 convolutional layer for channel fusion calculation and outputs the fused feature vector.

[0011] Optionally, S4 specifically includes: S41. Read the fused feature vector, treat each fused feature vector as an independent particle in a high-dimensional feature space, count the number of neighboring particles contained in the preset radius neighborhood of each particle, and use it as the density value of the particle. Calculate the Euclidean distance between each particle and all neighboring particles, and divide the particles with distance values ​​less than the preset density threshold into the same local high-density region. S42. In a local high-density region, calculate the distance product between any two particles, divide the dimension value of the fused feature vector by the distance product to obtain the gravitational value, and use the gravitational value to construct a gravitational field model. S43. Traverse all positions in the high-dimensional feature space, calculate the sum of gravitational potential energy from each position to all particles, generate a gravitational potential energy distribution surface, search for the minimum point on the gravitational potential energy distribution surface, and mark the minimum point as the center of the gravitational potential well. S44. Using the coordinates of each gravitational potential well center as the cluster center, generate dynamically adaptive label cluster centers, and assign all particles to the corresponding clusters based on the nearest gravitational potential well center, outputting the initial classification results.

[0012] Optionally, S5 specifically includes: S51. Read the initial classification results, create an empty classification decision set and an empty challenge verification set, store the coordinates of all label cluster centers in the initial classification results into the classification decision set, and select boundary samples within the preset radius of the label cluster centers into the challenge verification set. S52. In each round of calibration, read the fusion feature vector of the boundary samples in the challenge verification set, add random noise of a preset amplitude to the value of the fusion feature vector to generate perturbation data, and input the perturbation data into the classification decision set. S53. Calculate the Euclidean distance from each label cluster center in the classification decision set to the perturbation data, select the label cluster center with the smallest Euclidean distance as the classification response, and record the judgment result of whether the classification response label is consistent with the original label of the boundary sample. S54. Based on the judgment result, count the number of errors in the classification decision set response, divide the number of errors by the total number of samples in the challenge verification set to obtain the stability index of the current round, and determine whether the stability index is less than the preset convergence threshold or reaches the maximum number of rounds. If neither is met, return to the step of generating perturbation data to continue the next round of calibration. If any condition is met, read the final stability index. S55. Subtract the final stability index from the constant 1 to obtain the confidence weight. Multiply the confidence weight by the probability distribution of the initial classification result to obtain the corrected classification probability distribution. Use the confidence weight as the confidence score and output the corrected classification probability distribution and confidence score.

[0013] Optionally, S6 specifically includes: S61. Construct a first-in-first-out sliding window queue, store the values ​​in the corrected classification probability distribution into the sliding window queue in sequence, read the label prediction probability values ​​of all samples in the current sliding window queue, and calculate the arithmetic mean of the label prediction probability values ​​as the statistical mean of the current window. S62. Using the improved ADWIN algorithm, the sliding window queue is divided into a left sub-window and a right sub-window. The mean of the label prediction probability value of the left sub-window and the mean of the label prediction probability value of the right sub-window are calculated respectively. The absolute value of the difference between the mean of the left sub-window and the mean of the right sub-window is calculated as the distribution deviation value. S63. The number of data samples read from the left sub-window is recorded as the first sample number, and the number of data samples read from the right sub-window is recorded as the second sample number. The sum of the reciprocal of the first sample number and the reciprocal of the second sample number is calculated. The sum of the reciprocals is multiplied by the constant 4 to obtain the harmonic factor. The harmonic factor is divided by the sum of the first sample number and the second sample number to obtain the combined variance term. S64. Calculate the square of the distribution deviation value to obtain the deviation square term, read the preset reliability parameters of the improved ADWIN algorithm, construct the adaptive boundary calculation formula by using the product of the harmonic factor and the inverse natural logarithm function, calculate the weighted sum of the combined variance term, the deviation square term and the preset weight of the harmonic factor and the natural logarithm, and obtain the statistical distance. S65. Introduce the warning level of the DDM algorithm, set the warning threshold of the DDM algorithm and the drift judgment threshold of the improved ADWIN algorithm, and compare the statistical distance with the warning threshold set by the DDM algorithm. S66. When the statistical distance value exceeds the warning threshold set by DDM but does not exceed the drift judgment threshold set by the improved ADWIN algorithm, the pre-evolution mechanism is activated to lock the coordinate position of the existing nodes in the heterogeneous data association graph, read the new input data, calculate the Euclidean distance between the new input data and the nearest label cluster center, multiply the Euclidean distance value by a preset attenuation coefficient less than 1 to obtain the temporary connection weight, and replace the original connection weight with the temporary connection weight to reduce the update magnitude of the label cluster center by the new input data. S67. When the statistical distance value exceeds the drift judgment threshold set by the improved ADWIN algorithm, a dynamic evolution command is triggered to read the earliest stored historical data sample in the sliding window queue, calculate the Euclidean distance between the historical data sample and the current label cluster center, and delete the node corresponding to the historical data sample and the associated undirected edge from the heterogeneous data association graph when the Euclidean distance is greater than the preset forgetting radius. S68. Simultaneously read the fusion feature vector of the newly added input data and map it to a new node. Calculate the similarity between the new node and the remaining label cluster centers. Connect the new node to the label cluster center with the highest similarity. Recalculate and update the coordinate position of the label cluster center based on the feature vector value of the new node. Complete the incremental update of the heterogeneous data association map and label cluster centers.

[0014] Optionally, S7 specifically includes: S71. Read the preset multi-level tag library, input the text description information of each tag node in the multi-level tag library into the preset word vector mapping model to convert it into a high-dimensional tag semantic vector, and build a tag vector index table; S72. Read the heterogeneous data association graph after incremental update, traverse all node feature vectors in the heterogeneous data association graph, and calculate the cosine similarity between each node feature vector and all high-dimensional label semantic vectors in the label vector index table. S73. Select the label nodes with cosine similarity values ​​greater than the preset matching threshold as candidate nodes, sort the candidate nodes in descending order of cosine similarity values, generate a candidate node list, read the label cluster center after incremental update, calculate the path distance between the label cluster center to which the node feature vector belongs and each candidate node in the candidate node list, and select the node with the shortest path distance as the target node. S75. Starting from the target node, trace upwards to the root node of the multi-level tag library, record all parent node tags passed along the path in sequence, concatenate all parent node tags with the target node tags in hierarchical order to generate the optimal tag path, encapsulate the optimal tag path into a data structure of a preset format, generate the final structured tag classification result and output it.

[0015] The beneficial effects of this invention are: This invention addresses the challenges of large semantic gaps, difficulties in uncovering implicit associations, and limited feature extraction in multi-source heterogeneous data by constructing a heterogeneous data association graph and deploying a dual-stream coupled feature extraction network. It employs a unified high-dimensional feature space mapping and force-directed layout to construct node associations and optimize topology, generating fused feature vectors. A confidence calibration mechanism and an improved ADWIN algorithm are introduced to create a DDM algorithm early warning layer, performing disturbance resistance stability testing and distribution drift hierarchical response to output corrected classification results. In the dynamic evolution phase, a pre-evolution mechanism and incremental updates are implemented using statistical distance monitoring and dual threshold judgment to adaptively adjust the graph structure and cluster centers. Furthermore, an optimal label path matching mechanism completes the output of structured label classification results. Ultimately, this achieves closed-loop intelligent processing of deep semantic understanding, dynamic clustering evolution, and accurate label matching for multi-source heterogeneous big data, effectively improving classification accuracy in complex scenarios, model environmental adaptability, and long-term operational stability. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 This is a flowchart of a tag classification method based on big data analysis proposed in this invention; Fig. 2 This is a flowchart of the dual-stream coupled feature extraction network and feature fusion based on the bidirectional interactive fusion mechanism proposed in this invention. Fig. 3 This is a flowchart illustrating the data distribution drift monitoring and dynamic evolution process based on the DDM algorithm's early warning hierarchy and the improved ADWIN algorithm proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figs. 1-3 A tag classification method based on big data analysis includes the following steps: S1. Simultaneously collect multi-source heterogeneous raw big data, including text data and image / video data, and perform cleaning, noise reduction, and missing value completion to construct an initial sample pool. S2. Map the initial sample pool to a unified high-dimensional feature space, mine the implicit associations between data nodes, and construct a heterogeneous data association graph containing semantic topology and node feature vectors. S3. Input the heterogeneous data association graph into the dual-stream coupled feature extraction network. Extract the macro teacher flow with long-term semantic dependencies based on historical big data, and capture the micro student flow with instantaneous mutation features based on real-time data stream. Through a two-way interactive fusion mechanism, the macro teacher flow guides the micro student flow to focus on long-term value features. At the same time, the micro student flow feeds back the captured instantaneous abnormal features to the macro teacher flow for cognitive correction. Perform feature extraction and fusion on the heterogeneous data association graph, and output the fused feature vector. S4. Introduce the fused feature vector into the dynamic clustering algorithm based on the data gravitational field. Treat the data samples in the fused feature vector as spatial particles with mass and density, calculate the gravitational potential energy distribution in the high-dimensional feature space, search for the depth and center of the gravitational potential well, and generate dynamic adaptive label clustering centers and initial classification results. S5. Read the initial classification results, initialize a classification decision set and a challenge verification set. In each round of calibration, the challenge verification set generates perturbation data for the classification boundary in the classification decision set, and calculates the stability index based on the response of the classification decision set to the perturbation data. Repeat the above operations of generating perturbation data and calculating the stability index until the stability index reaches the preset convergence threshold or reaches the maximum number of rounds. Calculate the confidence weight based on the final stability index, use the confidence weight to perform weighted correction on the probability distribution of the initial classification results, and output the corrected classification probability distribution and confidence score. S6. Based on the corrected classification probability distribution and confidence score, the improved ADWIN algorithm is used to monitor the distribution characteristics in real time. The early warning level of the DDM algorithm is introduced. When the statistical distance exceeds the warning threshold set by the DDM algorithm, the pre-evolution mechanism is started. When the drift judgment threshold set by the improved ADWIN algorithm is exceeded, the dynamic evolution command is triggered to incrementally update the heterogeneous data association map and label cluster center. S7. Based on the incrementally updated heterogeneous data association map and label cluster centers, match the optimal label path from the preset multi-level label library, generate the final structured label classification result and output it.

[0019] This implementation significantly improves the accuracy, robustness, and real-time response capability of label classification. By constructing a heterogeneous data association graph containing semantic topology and node feature vectors, it achieves unified modeling and deep association of multi-source heterogeneous data such as text, images, and videos, effectively solving the problems of data silos and single feature representations. The dual-stream coupled feature extraction network, through bidirectional interaction between macroscopic teacher flow and microscopic student flow, not only utilizes historical data to capture long-term semantic dependencies but also sensitively responds to instantaneous changes in real-time data flow, ensuring the comprehensiveness and timeliness of feature extraction. A dynamic clustering algorithm based on data gravity field is introduced to simulate the physical potential field to search for gravitational potential wells, achieving adaptive generation of label cluster centers. The calibration mechanism of the challenge validation set evaluates boundary stability by generating perturbation data and uses confidence weighting to correct the probability distribution, significantly improving the credibility of the classification results. By combining the improved ADWIN and DDM algorithms for early warning, the system can monitor data distribution drift in real time and trigger dynamic evolution, incrementally updating the map and cluster centers to ensure that the model remains efficient and accurate even when the data environment changes drastically. This significantly enhances the system's intelligence level and application value in complex and dynamic scenarios.

[0020] In this embodiment, S1 specifically includes: S11. Simultaneously start the text data reading process and the image and video acquisition process through the preset synchronous triggering mechanism to collect multi-source heterogeneous raw big data, and respectively obtain the text data stream containing business records and the image and video data containing real-time scenes. S12. Scan each data record in the multi-source heterogeneous raw big data, use regular expression matching to identify non-standard characters and abnormal values, perform deletion operations on the identified non-standard characters and abnormal values, detect whether there are empty values ​​in the key fields of the data record, and when there are empty values, retrieve the statistical feature values ​​in the historical data and use the statistical feature values ​​to fill the empty values. S14. Map the text data and image / video data that have been cleaned, denoised, and had missing value filled into a unified storage structure for aggregation, and build an initial sample pool containing complete fields.

[0021] In this embodiment, S2 specifically includes: S21. Extract the pixel matrix of the image and video data in the initial sample pool and expand it into a numerical sequence. Use the bag-of-words model to convert the text data into a word frequency vector. Input the numerical sequence and word frequency vector into a preset fully connected layer for operation and map it into a vector space with a dimension of 128 to generate a node feature vector of uniform length. S22. Traverse all node feature vectors, multiply the corresponding dimension values ​​of any two node feature vectors and sum them to obtain the inner product, calculate the product of the moduli of the two node feature vectors, divide the inner product by the product to obtain the cosine value of the included angle, and when the cosine value of the included angle is greater than the preset similarity threshold, establish an undirected edge between the two nodes. S23. Starting from any node feature vector, traverse the network structure to obtain the neighbor node sets of two adjacent nodes and denote them as set A and set B respectively. Calculate the number of elements in the intersection of set A and set B and the number of elements in the union. Divide the number of elements in the intersection by the number of elements in the union to obtain the Jaccard coefficient. When the Jaccard coefficient is greater than a preset coefficient threshold, it is determined that there is a hidden association between the adjacent nodes. S24. Read the current horizontal and vertical coordinate values ​​of every two nodes in the heterogeneous data association map in the two-dimensional coordinate system, calculate the straight-line distance between the two nodes, calculate the repulsion force value by dividing the preset repulsion constant by the square of the straight-line distance, divide by the node mass to obtain the acceleration value, multiply the acceleration value by the preset time step to obtain the velocity value, and superimpose it on the current horizontal and vertical coordinates along the opposite direction of the line connecting the two nodes, and adjust the coordinate positions of the two nodes to make the nodes move away from each other. S25. Read the current horizontal and vertical coordinates of each pair of nodes connected by undirected edges in the heterogeneous data association graph in the two-dimensional coordinate system, calculate the straight-line distance between the two nodes, subtract the preset ideal side length distance to obtain the length difference, multiply the preset elastic coefficient by the length difference to calculate the tension value, divide by the node mass to obtain the acceleration value, multiply the acceleration value by the preset time step to obtain the velocity value, and superimpose it on the current horizontal and vertical coordinates along the direction of the line connecting the two nodes, and adjust the coordinate positions of the two nodes to make the nodes closer to each other. S26. The calculated repulsive force value and tensile force value are vector-superimposed in the horizontal and vertical coordinate directions to obtain the total displacement value. The total displacement value is added to the current horizontal and vertical coordinates of the node to obtain the updated coordinate position. The above calculation and update operation is repeated until the change in the node coordinate position is less than the set error threshold, and a heterogeneous data association graph containing semantic topology and node feature vectors is constructed.

[0022] In this embodiment, S3 specifically includes: S31. Input the heterogeneous data association map into the dual-stream coupled feature extraction network, read the data records in the initial sample pool, and divide the data records with timestamps earlier than the preset segmentation time into historical big data according to the timestamps of the data records, and divide the data records with timestamps later than the preset segmentation time into real-time data streams. S32. Configure a deep network structure containing three preset convolutional layers and two preset fully connected layers. Input historical big data into the macro teacher flow. In each convolutional layer, use preset convolutional kernels to perform sliding window convolution operations on the feature map of historical big data to extract long-term semantic dependency features in historical big data and generate macro feature maps. S33. Configure a lightweight network structure containing a Transformer encoder, read the image pixel matrix of the real-time data stream, and cut it into image blocks of a preset size in order from left to right and from top to bottom. Straighten the pixel values ​​in each image block into a one-dimensional column vector and arrange them in order to generate a vector sequence. Add the vector sequence to the preset learnable position encoding vector to obtain the position embedding vector. S34. Multiply the position embedding vector by three preset weight matrices to obtain the query vector, key vector, and value vector. Calculate the product of the query vector and the transpose of the key vector and divide by the square root of the vector dimension to obtain the attention score. Perform an exponential operation on the attention score and sum it row-wise to obtain the normalized denominator. Divide the attention score by the normalized denominator to obtain the autocorrelation weight matrix. Perform matrix multiplication on the autocorrelation weight matrix and the value vector to obtain the weighted feature vector. Pass the weighted feature vector through a preset feedforward neural network layer for linear transformation and nonlinear activation, and output a micro-feature map capturing transient change features. S35. Execute the two-way interactive fusion mechanism, read the average pixel value of each channel in the macro feature map, calculate the product of the average pixel value and the preset guidance coefficient, and perform an addition operation on the calculated product value and the real-time data stream. The macro feature map extracted from historical big data continues to guide the real-time data stream to focus on long-term value features. S36. Read feature points in the micro feature map whose gradient magnitude is greater than the preset mutation threshold, extract the position index and feature value of the feature points in the micro feature map, locate the corresponding position in the macro feature map corresponding to the historical big data according to the position index, add the feature value to the feature value of the corresponding position in the macro feature map, and continuously correct the macro feature map by the feedback of instantaneous abnormal features captured by the real-time data stream. S37. The macroscopic feature map and microscopic feature map after guided correction are concatenated and operated on in the channel dimension. The concatenated feature map is input into the final 1x1 convolutional layer for channel fusion calculation and outputs the fused feature vector.

[0023] This implementation innovatively designs a dual-stream coupled feature extraction architecture, which has significant advantages over single-model or simple feature concatenation methods. Traditional methods, if only using CNNs to extract historical features, often lack global semantic understanding; if only using Transformers to process real-time streams, although they can capture details, the computational burden is extremely heavy and they are easily affected by noise, making it difficult to take into account both long-term trends and instantaneous changes. This invention uses a macro-level teacher stream to efficiently extract long-term semantic dependencies from historical big data using convolutional neural networks, providing solid prior knowledge for classification; at the same time, it uses a lightweight Transformer encoder as a micro-level student stream to accurately capture instantaneous change features in the real-time data stream. In particular, the bidirectional interactive fusion mechanism in S35 and S36 guides the real-time stream to focus on long-term value through macro-level features and uses micro-level change features to correct macro-level cognition, achieving deep complementarity between historical and real-time information. This design effectively solves the limitations of single models in dynamic data environments while ensuring the integrity of feature representation, significantly improving the model's extraction accuracy and robustness for complex heterogeneous data features.

[0024] In this embodiment, S4 specifically includes: S41. Read the fused feature vector, treat each fused feature vector as an independent particle in a high-dimensional feature space, count the number of neighboring particles contained in the preset radius neighborhood of each particle, and use it as the density value of the particle. Calculate the Euclidean distance between each particle and all neighboring particles, and divide the particles with distance values ​​less than the preset density threshold into the same local high-density region. S42. In a local high-density region, calculate the distance product between any two particles, divide the dimension value of the fused feature vector by the distance product to obtain the gravitational value, and use the gravitational value to construct a gravitational field model. S43. Traverse all positions in the high-dimensional feature space, calculate the sum of gravitational potential energy from each position to all particles, generate a gravitational potential energy distribution surface, search for the minimum point on the gravitational potential energy distribution surface, and mark the minimum point as the center of the gravitational potential well. S44. Using the coordinates of each gravitational potential well center as the cluster center, generate dynamically adaptive label cluster centers, and assign all particles to the corresponding clusters based on the nearest gravitational potential well center, outputting the initial classification results.

[0025] In this embodiment, S5 specifically includes: S51. Read the initial classification results, create an empty classification decision set and an empty challenge verification set, store the coordinates of all label cluster centers in the initial classification results into the classification decision set, and select boundary samples within the preset radius of the label cluster centers into the challenge verification set. S52. In each round of calibration, read the fusion feature vector of the boundary samples in the challenge verification set, add random noise of a preset amplitude to the value of the fusion feature vector to generate perturbation data, and input the perturbation data into the classification decision set. S53. Calculate the Euclidean distance from each label cluster center in the classification decision set to the perturbation data, select the label cluster center with the smallest Euclidean distance as the classification response, and record the judgment result of whether the classification response label is consistent with the original label of the boundary sample. S54. Based on the judgment result, count the number of errors in the classification decision set response, divide the number of errors by the total number of samples in the challenge verification set to obtain the stability index of the current round, and determine whether the stability index is less than the preset convergence threshold or reaches the maximum number of rounds. If neither is met, return to the step of generating perturbation data to continue the next round of calibration. If any condition is met, read the final stability index. S55. Subtract the final stability index from the constant 1 to obtain the confidence weight. Multiply the confidence weight by the probability distribution of the initial classification result to obtain the corrected classification probability distribution. Use the confidence weight as the confidence score and output the corrected classification probability distribution and confidence score.

[0026] In this embodiment, S6 specifically includes: S61. Construct a first-in-first-out sliding window queue, store the values ​​in the corrected classification probability distribution into the sliding window queue in sequence, read the label prediction probability values ​​of all samples in the current sliding window queue, and calculate the arithmetic mean of the label prediction probability values ​​as the statistical mean of the current window. S62. Using the improved ADWIN algorithm, the sliding window queue is divided into a left sub-window and a right sub-window. The mean of the label prediction probability value of the left sub-window and the mean of the label prediction probability value of the right sub-window are calculated respectively. The absolute value of the difference between the mean of the left sub-window and the mean of the right sub-window is calculated as the distribution deviation value. S63. The number of data samples read from the left sub-window is recorded as the first sample number, and the number of data samples read from the right sub-window is recorded as the second sample number. The sum of the reciprocal of the first sample number and the reciprocal of the second sample number is calculated. The sum of the reciprocals is multiplied by the constant 4 to obtain the harmonic factor. The harmonic factor is divided by the sum of the first sample number and the second sample number to obtain the combined variance term. S64. Calculate the square of the distribution deviation value to obtain the deviation square term, read the preset reliability parameters of the improved ADWIN algorithm, construct the adaptive boundary calculation formula by using the product of the harmonic factor and the inverse natural logarithm function, calculate the weighted sum of the combined variance term, the deviation square term and the preset weight of the harmonic factor and the natural logarithm, and obtain the statistical distance. S65. Introduce the warning level of the DDM algorithm, set the warning threshold of the DDM algorithm and the drift judgment threshold of the improved ADWIN algorithm, and compare the statistical distance with the warning threshold set by the DDM algorithm. S66. When the statistical distance value exceeds the warning threshold set by the DDM algorithm but does not exceed the drift judgment threshold set by the improved ADWIN algorithm, the pre-evolution mechanism is activated to lock the coordinate position of the existing nodes in the heterogeneous data association graph, read the new input data, calculate the Euclidean distance between the new input data and the nearest label cluster center, multiply the Euclidean distance value by a preset attenuation coefficient less than 1 to obtain the temporary connection weight, and replace the original connection weight with the temporary connection weight to reduce the update magnitude of the label cluster center by the new input data. S67. When the statistical distance value exceeds the drift judgment threshold set by the improved ADWIN algorithm, a dynamic evolution command is triggered to read the earliest stored historical data sample in the sliding window queue, calculate the Euclidean distance between the historical data sample and the current label cluster center, and delete the node corresponding to the historical data sample and the associated undirected edge from the heterogeneous data association graph when the Euclidean distance is greater than the preset forgetting radius. S68. Simultaneously read the fusion feature vector of the newly added input data and map it to a new node. Calculate the similarity between the new node and the remaining label cluster centers. Connect the new node to the label cluster center with the highest similarity. Recalculate and update the coordinate position of the label cluster center based on the feature vector value of the new node. Complete the incremental update of the heterogeneous data association map and label cluster centers.

[0027] This invention achieves real-time distribution drift detection and incremental dynamic updates of the classification model by introducing an improved ADWIN algorithm combined with a DDM algorithm warning hierarchy. A sliding window queue is constructed to calculate the statistical mean and distribution deviation. The harmonic factor and adaptive boundary formula are used to accurately calculate the statistical distance, quantifying the degree of abrupt changes in data distribution. A tiered response is implemented when a risk signal is detected: exceeding the warning threshold triggers a pre-evolution mechanism to suppress interference from new data; exceeding the drift threshold triggers a dynamic evolution command, forgetting outdated historical nodes and updating cluster centers. This invention effectively addresses the concept drift problem in multi-source big data environments, maintaining long-term model stability while achieving agile adaptation to rapid changes in data patterns, significantly improving the long-term accuracy and robustness of label classification.

[0028] The improved ADWIN algorithm of this invention is similar to the original ADWIN algorithm in that both retain the core statistical monitoring architecture based on a sliding window queue. Specifically, it constructs a first-in-first-out sliding window queue to calculate the statistical mean of the current window. Using a segmentation technique, the queue is divided into a left sub-window and a right sub-window, and the difference between their means is calculated as the distribution deviation value. An adaptive boundary formula is constructed based on the harmonic factor, the reciprocal natural logarithm function, and preset confidence parameters. Finally, the relationship between the calculated statistical distance and a preset threshold is used to determine whether the data distribution has drifted.

[0029] The difference lies in that this invention breaks through the limitation of the original ADWIN algorithm, which relies solely on a single drift determination threshold for binary decision-making. It introduces the warning hierarchy of the DDM algorithm, constructing a dual-threshold hierarchical response system. While the original model only determines whether drift is triggered, this invention adds a numerical comparison step involving both the warning threshold and the drift determination threshold in step S65, subdividing the system state into three levels: stable, warning, and drift. Then, in step S66, when the statistical distance exceeds the warning threshold but not the drift threshold, a pre-evolution mechanism is initiated. This locks the node coordinates and uses a decay coefficient less than 1 to reduce the temporary connection weights of newly added data, suppressing interference from mutated data. Subsequently, in steps S67 and S68, a dynamic evolution command is triggered only when the drift determination threshold is confirmed to be exceeded. This involves calculating Euclidean distance to forget historical nodes and update the cluster centers.

[0030] The benefits of the improvements lie in the fact that, through the DDM algorithm's early warning hierarchy and dual-threshold response mechanism, the improved ADWIN algorithm can effectively distinguish between instantaneous noise interference in the data stream and real long-term concept drift, avoiding the resource waste and system oscillations caused by frequent model reconstruction when facing minor fluctuations in the original algorithm; the pre-evolution mechanism provides buffer protection in the early stages of drift, ensuring the continuity of classification; and the incremental update strategy ensures that when real drift occurs, the model can quickly eliminate outdated features and adaptively adjust cluster centers, significantly improving the anti-interference ability and long-term operational stability of the label classification system in complex data environments.

[0031] In this embodiment, S7 specifically includes: S71. Read the preset multi-level tag library, input the text description information of each tag node in the multi-level tag library into the preset word vector mapping model to convert it into a high-dimensional tag semantic vector, and build a tag vector index table; S72. Read the heterogeneous data association graph after incremental update, traverse all node feature vectors in the heterogeneous data association graph, and calculate the cosine similarity between each node feature vector and all high-dimensional label semantic vectors in the label vector index table. S73. Select the label nodes with cosine similarity values ​​greater than the preset matching threshold as candidate nodes, sort the candidate nodes in descending order of cosine similarity values, generate a candidate node list, read the label cluster center after incremental update, calculate the path distance between the label cluster center to which the node feature vector belongs and each candidate node in the candidate node list, and select the node with the shortest path distance as the target node. S75. Starting from the target node, trace upwards to the root node of the multi-level tag library, record all parent node tags passed along the path in sequence, concatenate all parent node tags with the target node tags in hierarchical order to generate the optimal tag path, encapsulate the optimal tag path into a data structure of a preset format, generate the final structured tag classification result and output it.

[0032] Example 1: To verify the practical feasibility and technical superiority of this invention in the field of large-scale internet content ecosystem governance, this invention was deployed in the intelligent review and distribution center of a leading short video and social content platform. This center handles hundreds of millions of dynamic content processing tasks daily on the platform. This scenario covers more than 500 vertical content areas, including entertainment, technology, education, and financial information, with more than 20 million new user-generated content (UGC) items added daily, including videos, bullet comments, reviews, article titles, and hashtags. Before applying this invention, the existing content tag classification system was mainly based on traditional static text matching and shallow deep learning models. When faced with massive, noisy, and semantically variable unstructured content data, it often encountered problems such as biased understanding of tag semantics, slow cold start of hot topics, and lagging recognition of popular internet slang. Especially in scenarios with sudden traffic surges, such as large-scale live events and sudden social hot topics, the system struggled to capture the drift of semantic distribution in real time, resulting in decreased recommendation matching accuracy, missed violations, or insufficient exposure of high-quality content, failing to meet the stringent requirements of the content ecosystem for accurate distribution and security control.

[0033] In this complex online content ecosystem, the method of this invention is deeply integrated into the core links of content understanding and tag generation, aiming to solve the problem of unifying semantic representation and dynamic evolution of multimodal content data. The system first performs multidimensional alignment and cleaning on extracted video frame keyframe features, speech recognition-transcribed text, bullet comment semantic vectors, and user interaction behavior data to construct a heterogeneous content semantic association graph. Based on this, the dual-stream coupled feature extraction network employed in this invention plays a crucial role. The network processes historically accumulated content data and real-time incoming traffic data in parallel. The macro-stream focuses on extracting long-term stable domain common features and content category boundaries, while the micro-stream uses a lightweight encoder to keenly capture real-time online memes, breaking news keywords, and instantaneous high-interaction features. Through a two-way interactive fusion mechanism, the historical knowledge system can effectively guide the focus of real-time features, avoiding being misled by short-term noise. Simultaneously, the emerging semantic features captured in real-time can provide feedback to correct macro-category boundaries, thereby generating a high-dimensional fused feature vector that comprehensively reflects the deep meaning of the content. To address the semantic drift caused by the rapid emergence of new words, phrases, and trending topics on the internet, the system utilizes an improved ADWIN algorithm with a DDM algorithm-based early warning system to monitor content tag distribution in real time. Once a significant shift in tag distribution characteristics within a specific domain is detected, the algorithm immediately triggers an early warning and initiates dynamic updates to model parameters, ensuring that the tag system remains highly aligned with current online trends. Table 1 below shows a comparison of tag classification and semantic awareness performance between the proposed method and traditional methods over a two-month period. Table 1. Detailed comparison of label classification and semantic perception performance between the present invention and traditional methods.

[0034] Based on the comparative data shown in Table 1, it can be seen that the tag classification method based on big data analysis proposed in this invention shows significant performance advantages over traditional static methods in the field of multimodal Internet content processing, especially in key indicators such as classification accuracy, processing timeliness, drift response speed, and mislabeling control.

[0035] In terms of tag classification accuracy, this invention maintains a high level of over 92% in all five typical content scenarios, far exceeding the average accuracy of traditional systems (approximately 78%). For example, in the scenario of "internet meme outbreak," traditional methods, limited by fixed semantic models, only achieve an accuracy of 65.4% when faced with new words; while this invention, through a dual-stream coupling mechanism to capture real-time semantic features, achieves an accuracy of 92.1%, effectively solving the problem of identifying sudden semantic events.

[0036] In terms of handling timeliness and drift response, this invention controls the average feature extraction time to within 25 milliseconds and shortens the semantic drift response latency to within 8 seconds. Especially in high-concurrency drift scenarios such as "live broadcasts of breaking events," traditional methods require 180.5 seconds to respond, while this invention only requires 8.2 seconds, ensuring agile capture of trending topics.

[0037] Furthermore, this invention controls the mislabeling rate to below 4.2%, significantly outperforming traditional methods. Overall, the performance improvement of this invention is particularly significant when dealing with high-noise, dynamic, and sudden content, making it highly valuable for practical applications.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A tag classification method based on big data analysis, characterized in that, Includes the following steps: S1. Simultaneously collect raw big data from multiple heterogeneous sources, and perform cleaning, noise reduction, and missing value completion to construct an initial sample pool; S2. Map the initial sample pool to a unified high-dimensional feature space, mine the implicit associations between data nodes, and construct a heterogeneous data association graph; S3. Input the heterogeneous data association map into the dual-stream coupled feature extraction network to extract the macro-level teacher flow with long-term semantic dependence and capture the micro-level student flow with transient mutation features. Through a bidirectional interactive fusion mechanism, the heterogeneous data association map is used to extract and fuse features and output a fused feature vector. S4. Introduce the fused feature vector into the dynamic clustering algorithm based on the data gravitational field. Treat the data samples as spatial particles with mass and density, calculate the gravitational potential energy distribution in the high-dimensional feature space, search for the depth and center of the gravitational potential well, and generate the label clustering center and the initial classification result. S5. Read the initial classification results, initialize a classification decision set and a challenge verification set. In each round of calibration, the challenge verification set generates perturbation data and calculates the stability index until the stability index reaches the preset convergence threshold or the maximum number of rounds is reached. Calculate the confidence weight, perform weighted correction, and output the corrected classification probability distribution and confidence score. S6. Based on the corrected classification probability distribution and confidence score, the improved ADWIN algorithm is used to monitor the distribution characteristics in real time. The early warning level of the DDM algorithm is introduced. When the statistical distance exceeds the warning threshold set by the DDM algorithm, the pre-evolution mechanism is started. When the drift judgment threshold set by the improved ADWIN algorithm is exceeded, the dynamic evolution command is triggered to incrementally update the heterogeneous data association map and label cluster center. S7. Based on the incrementally updated heterogeneous data association map and label cluster centers, match the optimal label path from the preset multi-level label library, generate the final structured label classification result and output it.

2. The tag classification method based on big data analysis according to claim 1, characterized in that, S1 specifically includes: S11. Simultaneously start the text data reading process and the image and video acquisition process through the preset synchronous triggering mechanism to collect multi-source heterogeneous raw big data, and respectively obtain the text data stream containing business records and the image and video data containing real-time scenes. S12. Scan each data record in the multi-source heterogeneous raw big data, use regular expression matching to identify non-standard characters and abnormal values, perform deletion operations on the identified non-standard characters and abnormal values, detect whether there are empty values ​​in the key fields of the data record, and when there are empty values, retrieve the statistical feature values ​​in the historical data and use the statistical feature values ​​to fill the empty values. S14. Map the text data and image / video data that have been cleaned, denoised, and had missing value imputation operations to a unified storage structure for aggregation, and construct an initial sample pool.

3. The tag classification method based on big data analysis according to claim 1, characterized in that, S2 specifically includes: S21. Extract the pixel matrix of the image and video data in the initial sample pool and expand it into a numerical sequence. Use the bag-of-words model to convert the text data into word frequency vectors. Input the numerical sequence and word frequency vectors into a preset fully connected layer for operation and map them into a vector space with a dimension of 128 to generate node feature vectors. S22. Traverse all node feature vectors, multiply the corresponding dimension values ​​of any two node feature vectors and sum them to obtain the inner product, calculate the product of the moduli of the two node feature vectors, divide the inner product by the product to obtain the cosine value of the included angle, and when the cosine value of the included angle is greater than the preset similarity threshold, establish an undirected edge between the two nodes. S23. Starting from any node feature vector, traverse the network structure to obtain the neighbor node sets of two adjacent nodes and denote them as set A and set B respectively. Calculate the number of elements in the intersection of set A and set B and the number of elements in the union. Divide the number of elements in the intersection by the number of elements in the union to obtain the Jaccard coefficient. When the Jaccard coefficient is greater than a preset coefficient threshold, it is determined that there is a hidden association between the adjacent nodes. S24. Read the current horizontal and vertical coordinate values ​​of every two nodes in the heterogeneous data association map in the two-dimensional coordinate system, calculate the straight-line distance between the two nodes, calculate the repulsion force value by dividing the preset repulsion constant by the square of the straight-line distance, divide by the node mass to obtain the acceleration value, multiply the acceleration value by the preset time step to obtain the velocity value, and superimpose it on the current horizontal and vertical coordinates along the opposite direction of the line connecting the two nodes, and adjust the coordinate positions of the two nodes to make the nodes move away from each other. S25. Read the current horizontal and vertical coordinates of each pair of nodes connected by undirected edges in the heterogeneous data association graph in the two-dimensional coordinate system, calculate the straight-line distance between the two nodes, subtract the preset ideal side length distance to obtain the length difference, multiply the preset elastic coefficient by the length difference to calculate the tension value, divide by the node mass to obtain the acceleration value, multiply the acceleration value by the preset time step to obtain the velocity value, and superimpose it on the current horizontal and vertical coordinates along the direction of the line connecting the two nodes, and adjust the coordinate positions of the two nodes to make the nodes closer to each other. S26. The calculated repulsive force value and tensile force value are vector-superimposed in the horizontal and vertical coordinate directions to obtain the total displacement value. The total displacement value is added to the current horizontal and vertical coordinates of the node to obtain the updated coordinate position. The above calculation and update operation is repeated until the change in the node coordinate position is less than the set error threshold, and a heterogeneous data association map is constructed.

4. The tag classification method based on big data analysis according to claim 1, characterized in that, S3 specifically includes: S31. Input the heterogeneous data association map into the dual-stream coupled feature extraction network, read the data records in the initial sample pool, and divide the data records with timestamps earlier than the preset segmentation time into historical big data according to the timestamps of the data records, and divide the data records with timestamps later than the preset segmentation time into real-time data streams. S32. Configure a deep network structure containing three preset convolutional layers and two preset fully connected layers. Input historical big data into the macro teacher flow. In each convolutional layer, use preset convolutional kernels to perform sliding window convolution operations on the feature map of historical big data to extract long-term semantic dependency features in historical big data and generate macro feature maps. S33. Configure a lightweight network structure containing a Transformer encoder, read the image pixel matrix of the real-time data stream, and cut it into image blocks of a preset size in order from left to right and from top to bottom. Straighten the pixel values ​​in each image block into a one-dimensional column vector and arrange them in order to generate a vector sequence. Add the vector sequence to the preset learnable position encoding vector to obtain the position embedding vector. S34. Multiply the position embedding vector by three preset weight matrices to obtain the query vector, key vector, and value vector. Calculate the product of the query vector and the transpose of the key vector and divide by the square root of the vector dimension to obtain the attention score. Perform an exponential operation on the attention score and sum it row-wise to obtain the normalized denominator. Divide the attention score by the normalized denominator to obtain the autocorrelation weight matrix. Perform matrix multiplication on the autocorrelation weight matrix and the value vector to obtain the weighted feature vector. Pass the weighted feature vector through a preset feedforward neural network layer for linear transformation and nonlinear activation to output the micro-feature map. S35. Execute the two-way interactive fusion mechanism, read the average pixel value of each channel in the macro feature map, calculate the product of the average pixel value and the preset guidance coefficient, and perform an addition operation on the calculated product value and the real-time data stream. The macro feature map extracted from historical big data continues to guide the real-time data stream to focus on long-term value features. S36. Read feature points in the micro feature map whose gradient magnitude is greater than the preset mutation threshold, extract the position index and feature value of the feature points in the micro feature map, locate the corresponding position in the macro feature map corresponding to the historical big data according to the position index, add the feature value to the feature value of the corresponding position in the macro feature map, and continuously correct the macro feature map by the feedback of instantaneous abnormal features captured by the real-time data stream. S37. The macroscopic feature map and microscopic feature map after guided correction are concatenated and operated on in the channel dimension. The concatenated feature map is input into the final 1x1 convolutional layer for channel fusion calculation and outputs the fused feature vector.

5. The tag classification method based on big data analysis according to claim 1, characterized in that, S4 specifically includes: S41. Read the fused feature vector, treat each fused feature vector as an independent particle in a high-dimensional feature space, count the number of neighboring particles contained in the preset radius neighborhood of each particle, and use it as the density value of the particle. Calculate the Euclidean distance between each particle and all neighboring particles, and divide the particles with distance values ​​less than the preset density threshold into the same local high-density region. S42. In a local high-density region, calculate the distance product between any two particles, divide the dimension value of the fused feature vector by the distance product to obtain the gravitational value, and use the gravitational value to construct a gravitational field model. S43. Traverse all positions in the high-dimensional feature space, calculate the sum of gravitational potential energy from each position to all particles, generate a gravitational potential energy distribution surface, search for the minimum point on the gravitational potential energy distribution surface, and mark the minimum point as the center of the gravitational potential well. S44. Using the coordinates of each gravitational potential well center as the cluster center, generate dynamically adaptive label cluster centers, and assign all particles to the corresponding clusters based on the nearest gravitational potential well center, outputting the initial classification results.

6. The tag classification method based on big data analysis according to claim 1, characterized in that, S5 specifically includes: S51. Read the initial classification results, create an empty classification decision set and an empty challenge verification set, store the coordinates of all label cluster centers in the initial classification results into the classification decision set, and select boundary samples within the preset radius of the label cluster centers into the challenge verification set. S52. In each round of calibration, read the fusion feature vector of the boundary samples in the challenge verification set, add random noise of a preset amplitude to the value of the fusion feature vector to generate perturbation data, and input the perturbation data into the classification decision set. S53. Calculate the Euclidean distance from each label cluster center in the classification decision set to the perturbation data, select the label cluster center with the smallest Euclidean distance as the classification response, and record the judgment result of whether the classification response label is consistent with the original label of the boundary sample. S54. Based on the judgment result, count the number of errors in the classification decision set response, divide the number of errors by the total number of samples in the challenge verification set to obtain the stability index of the current round, and determine whether the stability index is less than the preset convergence threshold or reaches the maximum number of rounds. If neither is met, return to the step of generating perturbation data to continue the next round of calibration. If any condition is met, read the final stability index. S55. Subtract the final stability index from the constant 1 to obtain the confidence weight. Multiply the confidence weight by the probability distribution of the initial classification result to obtain the corrected classification probability distribution. Use the confidence weight as the confidence score and output the corrected classification probability distribution and confidence score.

7. The tag classification method based on big data analysis according to claim 1, characterized in that, S6 specifically includes: S61. Construct a first-in-first-out sliding window queue, store the values ​​in the corrected classification probability distribution into the sliding window queue in sequence, read the label prediction probability values ​​of all samples in the current sliding window queue, and calculate the arithmetic mean of the label prediction probability values ​​as the statistical mean of the current window. S62. Using the improved ADWIN algorithm, the sliding window queue is divided into a left sub-window and a right sub-window. The mean of the label prediction probability value of the left sub-window and the mean of the label prediction probability value of the right sub-window are calculated respectively. The absolute value of the difference between the mean of the left sub-window and the mean of the right sub-window is calculated as the distribution deviation value. S63. The number of data samples read from the left sub-window is recorded as the first sample number, and the number of data samples read from the right sub-window is recorded as the second sample number. The sum of the reciprocal of the first sample number and the reciprocal of the second sample number is calculated. The sum of the reciprocals is multiplied by the constant 4 to obtain the harmonic factor. The harmonic factor is divided by the sum of the first sample number and the second sample number to obtain the combined variance term. S64. Calculate the square of the distribution deviation value to obtain the deviation square term, read the preset reliability parameters of the improved ADWIN algorithm, construct the adaptive boundary calculation formula by using the product of the harmonic factor and the inverse natural logarithm function, calculate the weighted sum of the combined variance term, the deviation square term and the preset weight of the harmonic factor and the natural logarithm, and obtain the statistical distance. S65. Introduce the warning level of the DDM algorithm, set the warning threshold of the DDM algorithm and the drift judgment threshold of the improved ADWIN algorithm, and compare the statistical distance with the warning threshold set by the DDM algorithm. S66. When the statistical distance value exceeds the warning threshold set by the DDM algorithm but does not exceed the drift judgment threshold set by the improved ADWIN algorithm, the pre-evolution mechanism is activated to lock the coordinate position of the existing nodes in the heterogeneous data association graph, read the new input data, calculate the Euclidean distance between the new input data and the nearest label cluster center, multiply the Euclidean distance value by a preset attenuation coefficient less than 1 to obtain the temporary connection weight, and replace the original connection weight with the temporary connection weight to reduce the update magnitude of the label cluster center by the new input data. S67. When the statistical distance value exceeds the drift judgment threshold set by the improved ADWIN algorithm, a dynamic evolution command is triggered to read the earliest stored historical data sample in the sliding window queue, calculate the Euclidean distance between the historical data sample and the current label cluster center, and delete the node corresponding to the historical data sample and the associated undirected edge from the heterogeneous data association graph when the Euclidean distance is greater than the preset forgetting radius. S68. Simultaneously read the fusion feature vector of the newly added input data and map it to a new node. Calculate the similarity between the new node and the remaining label cluster centers. Connect the new node to the label cluster center with the highest similarity. Recalculate and update the coordinate position of the label cluster center based on the feature vector value of the new node. Complete the incremental update of the heterogeneous data association map and label cluster centers.

8. The tag classification method based on big data analysis according to claim 1, characterized in that, Specifically, S7 includes: S71. Read the preset multi-level tag library, input the text description information of each tag node in the multi-level tag library into the preset word vector mapping model to convert it into a high-dimensional tag semantic vector, and build a tag vector index table; S72. Read the heterogeneous data association graph after incremental update, traverse all node feature vectors in the heterogeneous data association graph, and calculate the cosine similarity between each node feature vector and all high-dimensional label semantic vectors in the label vector index table. S73. Select the label nodes with cosine similarity values ​​greater than the preset matching threshold as candidate nodes, sort the candidate nodes in descending order of cosine similarity values, generate a candidate node list, read the label cluster center after incremental update, calculate the path distance between the label cluster center to which the node feature vector belongs and each candidate node in the candidate node list, and select the node with the shortest path distance as the target node. S75. Starting from the target node, trace upwards to the root node of the multi-level tag library, record all parent node tags passed along the path in sequence, concatenate all parent node tags with the target node tags in hierarchical order to generate the optimal tag path, encapsulate the optimal tag path into a data structure of a preset format, generate the final structured tag classification result and output it.

Images